- Load offline workloads, prune metrics, pre-process knobs.
- Saves pruned metrics, knob encoders to be used later to transform online workloads.
- Create offline Workload objects from processed offline data from DataRepo
- Given online workload and test knobs, performs workload mapping
- Augments current workload with mapped workload
- Predict latency for test knobs
- Uses pruned metric info and trained knob encoders in DataRepo to transform online workloads
- Models a single workload
- Contains GPR models trained on each metric (knobs -> GPR -> latency/metric)
- Works given a mode ('val' or 'test')
- Loads online workloads and their respective test knobs
- If 'val', loads true latencies to report mean absolute error (MAE)
- Saves a result file under
./data/out/with mapped workload info and predicted latency - Result file name
val_results_({MAE}).csv
- Train - offline_workload.csv (makes our DataRepo)
- Val – online_workload_B.csv (100 workloads with 6 configs each) split into 3 files
online_workload.csv(100 workloads with 5 configs each, randomly chosen)test_knobs.csv(100 workloads with left out 1 config each)true_latency.csv(True latency values for each test knob in test_knobs.csv)
- Test – online_workload_C.csv and provided test knobs